{
  "schema": "experiment-results/v0",
  "experiment_id": "EXP-002",
  "protocol_freeze": "lab/exp/exp-002-protocol.json",
  "executed_at": "2026-04-28",
  "status": "confirmed",
  "n_runs": 1437,
  "n_regressions": 960,
  "domains": ["refund", "claim", "memory", "code"],

  "primary_result": {
    "metric": "theta_OPBR",
    "value": 0.914,
    "wilson_95_ci": [0.894, 0.932],
    "n_regressions": 960
  },

  "detector_F1": {
    "behavior_contract":           0.768,
    "rich_output_provenance":      0.890,
    "llm_judge_trajectory":        0.512,
    "output_only_field_check":     0.086,
    "operational_composite":       0.094,
    "output_plus_operational":     0.131
  },

  "preregistered_gates": {
    "G_theta_floor":        { "threshold": 0.80, "observed": 0.914, "passes": true },
    "G_wilson_lower_floor": { "threshold": 0.75, "observed": 0.894, "passes": true }
  },

  "verdict": "Both gates pass. Lower Wilson bound exceeds the preregistered floor by 14.4 percentage points. theta_OPBR = 91.4% — under non-instructed perturbations across four domains, nearly every behavioral regression also produces a passing output.",

  "regime_inversion_finding": {
    "best_detector_under_perturbation": "rich_output_provenance",
    "F1": 0.890,
    "mechanism": "Under perturbations like latency pressure and prompt compression, an agent that skips a precondition also fills fewer provenance fields and produces sparser outputs. The rich-output detector catches the correlated output degradation, not the behavioral failure directly. Its F1 = 0.890 is real, but the mechanism is wrong.",
    "why_this_matters": "The Study 2 vs. Study 5 (EXP-004) inversion is the load-bearing finding for the paper's baseline-selection discussion. Detector ranking is regime-dependent; on canonical OPBR (EXP-004) the same baseline drops to F1 = 0.400 while contracts reach F1 = 0.982.",
    "recall_gap_explanation": "Contract recall is 0.786 (gap of 0.214); enumerable specification work — see EXP-005 completeness analyzer."
  },

  "replication": {
    "cmd": "abw run --experiment EXP-002",
    "code": "https://github.com/raising-agents/agent-behavior-workbench",
    "traces_corpus": "OPBR-Bench v0"
  },

  "cite_as": "Sanchez de la Sierra, A. (2026). EXP-002 — Semi-blind perturbation. Raising Agents Lab. https://raisingagents.is/lab/exp/exp-002"
}